Recently, overlay networks have been proposed to address the problem of scalability in providing video-on-demand (VoD) service. However, from the perspective of service providing, their efficiency has not been carefully studied and still remains far from clear, especially considering the impacts of user interactivities and in the case of multiple files with different and varying popularities on sharing. Towards this end, in this paper, by analyzing more than 20,000,000 real workload traces, we first identify two practical factors which we believe have determinant impacts on the scalability: user interactivities and popularity differences among files. Then we further evaluate cache-and-relay (CR), a representative scheme of overlay networks, with the real workload traces. Simulation results show that CR only save about half of the server bandwidth even when there is no buffer constraint at clients, not so scalable as our original expectation.